num_frames = 1000000 # total frames that will be learning from
batch_size = 32 # the number of samples that are provided to the model for update services at a given time
gamma = 0.99 # the discount of future rewards
record_idx = 10000 #
replay_initial = 10000 # number frames that are held
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(
torch.load("model_pretrained.pth",
map_location='cpu')) #loading in the pretrained model
target_model = QLearner(env, num_frames, batch_size, gamma,
replay_buffer) #load in model
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(),
lr=0.0001) #learning rate set and optimizing the model
if USE_CUDA:
model = model.cuda() # sends model to gpu
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000 #used in ?
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
# .01 + 0.99* 1/ e^(frame index / 30000)
# So epsilon starts at pretty much 1
# as frame index increases, exp will get larger so 1/exp will decrease, so the 0.99 term will decrease, leaving us with just the final
USE_CUDA = torch.cuda.is_available()
# Set up game
env_id = "PongNoFrameskip-v4"
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
replay_buffer = ReplayBuffer(replay_buff_size) # Buffer size
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # Create model
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
model.eval()
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer) # Create target model
target_model.copy_from(model)
# Optimize model's parameters
optimizer = optim.Adam(model.parameters(), lr=lr)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
# Neg exp func. Start exploring then exploiting according to frame_indx
epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
episode_reward = 0
state = env.reset() # Initial state
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("model_pretrained.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model)
optimizer = optim.Adam(model.parameters(), lr=0.00001)
if USE_CUDA:
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
env = make_atari(env_id)
env = wrap_deepmind(env)
env = wrap_pytorch(env)
num_frames = 1000000
batch_size = 32 #num samples provided to model at a time to update
gamma = 0.99
record_idx = 10000
replay_initial = 10000
replay_buffer = ReplayBuffer(100000)
model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
model.load_state_dict(torch.load("modelsave.pth", map_location='cpu'))
target_model = QLearner(env, num_frames, batch_size, gamma, replay_buffer)
target_model.copy_from(model) #prevents too many radical changes
optimizer = optim.Adam(model.parameters(), lr=0.0001)
if USE_CUDA: #GPU stuff
model = model.cuda()
target_model = target_model.cuda()
print("Using cuda")
epsilon_start = 0.5 # Q-learning "error"
epsilon_final = 0.01
epsilon_decay = 30000
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
losses = []
all_rewards = []
